Custom Instructions

Customize how AI processes your transcriptions.

The free plan includes 1 Tone with its own Custom Instructions field, so you can try this Pro feature before upgrading.

Overview

Custom Instructions let you guide how the AI handles your text. They’re added to the system prompt used during AI Enhancement and Command Mode.

Where to Set Instructions

Two scopes, set in different places.

Global

Settings > AI Setup > Custom Instructions

Applies to every transcription regardless of which app or mode you’re using. Acts as the baseline guidance for all dictation.

Per-Tone

Settings > Tones > [Tone Name] > Custom Instructions

Applies only when that Tone is active, based on the focused app. See the Tones page for setup and examples.

When a Tone is active, both stacks are sent to the AI together. Global instructions provide baseline guidance and the Tone’s instructions add app-specific directives on top.

What to Include

Formatting Preferences

• “Use bullet points for lists”
• “Use Oxford commas”
• “Keep paragraphs to 2-3 sentences”
• “Format dates as DD/MM/YYYY”

Name Corrections

• “My name is spelled ‘Ijuel’ not ‘EOL’”
• “The product is ‘PostgreSQL’ not ‘post-gress’”
• “Refer to my company as ‘Acme Corp’ not ‘acmee’ or ‘akmay’”

Style Directives

• “Write in active voice”
• “Avoid exclamation marks”
• “Use American English spelling”
• “Keep sentences under 20 words”

Content Rules

• “Never add greetings or sign-offs”
• “Don’t convert technical terms to plain English”
• “Preserve code-related terms exactly as spoken”

Examples in action

Date and number formatting

Instructions: “Format dates as DD MMM YYYY (e.g., ‘12 May 2026’). Spell out numbers one to nine; use digits for 10 and above. Add a comma in numbers ≥ 1,000.”

How it helps: Locks every transcription to a consistent number and date style without you reformatting manually, which is especially useful for reports, status updates, and metrics summaries.

Default output

We launched on May 12 2026 and got 2500 signups in the first 3 days. 3 of the integrations broke and there were 15 customer complaints.

It writes

We launched on 12 May 2026 and got 2,500 signups in the first three days. Three of the integrations broke and there were 15 customer complaints.

British English spelling

Instructions: “Use British English spelling throughout: colour, organisation, realise, favour, behaviour, centre, theatre, programme. Maintain British conventions everywhere, including in word endings (‘-ise’ not ‘-ize’).”

How it helps: Locks every transcription to British conventions without you having to manually fix US spellings each time.

Default output

The organization’s color scheme didn’t favor the team’s behavior; we’ll need to realize a new design centered on user feedback.

It writes

The organisation’s colour scheme didn’t favour the team’s behaviour; we’ll need to realise a new design centred on user feedback.

Brand and jargon preservation

Instructions: “Always spell brand and product names correctly: PostgreSQL (not ‘post-gress’ or ‘postgres q l’), Anthropic (not ‘antropic’), TypeScript (not ‘type script’), Cloudflare (one word, capital C). Preserve exact technical terms as spoken.”

How it helps: Stops the AI from “correcting” technical vocabulary it doesn’t recognise, so brand names and product names land right every time.

You speak

We’re migrating from post-gress to a new antropic-backed system, with type script and cloud flare in front.

It writes

We’re migrating from PostgreSQL to a new Anthropic-backed system, with TypeScript and Cloudflare in front.

Active voice rewrite

Instructions: “Convert passive voice to active voice. Identify the actor and restructure the sentence so they’re the subject. Keep the original meaning intact.”

How it helps: Makes spoken thoughts read more direct and decisive on the page, which is especially useful for status updates, reports, and PRs.

You speak

The bug was identified by the QA team after the deployment was completed by engineering, and a fix has been proposed by the senior developer.

It writes

The QA team identified the bug after engineering completed the deployment, and the senior developer has proposed a fix.

Tips

• Be specific: “use bullet points” works better than “format nicely”
• Test your instructions with a few transcriptions and adjust
• Don’t over-constrain: too many rules can confuse smaller AI models
• Instructions work best with models that are good at following directions (GPT-4, Claude, Gemini)

Set up an AI provider

Connect Groq, OpenAI, Anthropic, Gemini, or any of 10+ supported providers.